A fine-tune of Qwen/Qwen3.5-9B (reinforcement learning, 50 optimizer steps). Weights, config
and tokenizer as saved by the trainer; load with AutoModelForCausalLM.from_pretrained.
Step 50 of up_cmh_cispo_q359b_judge2_k4_16k_v5: Qwen3.5-9B, teams of 4 by mail, LLM-judge credit assignment (gemini-3.8-flash, rubric v2), CISPO 16k on contest_math_hard_v2 (wandb a7855255).
Qwen/Qwen3.5-9B